Is your Claude / GPT real?
Detect whether your API relay's upstream has been swapped for a cheaper model (e.g. Claude → Qwen / DeepSeek / GLM / Doubao / Kimi and other Chinese models). Behavioral fingerprinting — it sees through upstreams prompted to claim they are Claude.
Our server uses your downstream token to send a set of carefully designed probe questions to the relay, just like a normal user.
We compare the answers' style fingerprint, self-identification and capability tells — even an upstream prompted to claim it is Claude won't match the real fingerprint.
You get a Genuine / Suspected-substitution / Inconclusive verdict + evidence; edge cases can be settled by cryptographic deep verification (a relay can't fake it without the upstream's signature).
| Method | Key needed | Speed | Accuracy | Best for |
|---|---|---|---|---|
| B Web quick-check | No (paste output) | Fastest | Screening | A quick first look |
| A Local CLI recommended | Yes (local, never uploaded) | Medium | Most accurate | A reliable verdict |
| C Free deep verify | Yes (this probe only, not stored) | Slower | Cryptographic proof (Claude only) | Settling edge cases |
Method A — Local CLI · most accurate · key never leaves your machine
RecommendedDownload one file and run it — it will ask you for the details interactively. Your key is used only to call your own endpoint, never uploaded.
curl -O https://panshi.io/relay-check/relay_check.py && python3 relay_check.py
Python 3 only, no dependencies. It prompts for your relay URL, key, and model, then prints Genuine / Suspected-substitution / Inconclusive + evidence.
Method B — Quick web check · no key needed · paste output
Pick a specific tier to catch "intra-vendor downgrade" (e.g. labeled Opus, served Haiku).
⚠️ Results are probabilistic signals, not legal proof. Quantized / distilled variants and models outside our reference set cannot be determined. Detection by panshi.io.
Method C — free deep verification · cryptographic signature · tamper-evident
3/3 free left todayFingerprint says "suspicious but unsure"? This settles it with cryptographic verification that a relay can't fake without the upstream's signature — Gemini/DeepSeek impersonators fail. Claude models only. Key is used for this active probe and not stored. Privacy-conscious? Use Method A (local CLI).
We keep upgrading detection. Detection details are not disclosed (to prevent evasion).
- v0.8 2026-06-19Hard intra-vendor downgrade detection — "pay for Opus, get Haiku" now reliably flagged (FP-validated), not just a soft hint.
- v0.7 2026-06-19Anti-persona-injection hardening — for high-value Claude/Gemini claims, web-only fingerprint can be skewed by a strong impersonation persona; now auto-lowers confidence and strongly pushes cryptographic deep verify (Method C) instead of declaring "genuine".
- v0.6 2026-06-19Chinese-model detection added — GLM (Zhipu) / Doubao (ByteDance) / MiniMax and other common cheap backends; catches relays that swap Claude/GPT for these.
- v0.5 2026-06-18Intra-vendor downgrade detection (e.g. labeled Opus, served a cheaper tier) + steadier, harder to evade.
- v0.4 2026-06-18Stronger multi-signal detection (harder to fool with disguise prompts); advanced disguises prompt cryptographic deep verification; graceful degradation when the engine is down; result feedback.
- v0.3 2026-06-18Anti-disguise detection — catches cheap-model backends prompted to impersonate Claude/Gemini.
- v0.2 2026-06-17Centered-fingerprint rebuild, no false positives observed on official endpoints in our testing.
- v0.1Behavioral-fingerprint detection tool launched.
Relays swap models silently — what is genuine today can be quantized tomorrow. Drop your email; we will tell you when relay-substitution risk shifts.